AtlasSphere

This project aims to sample a batch of training paths for MPNet. It consists of two parts: C++ part and Python part. The two part are connected by Pybind11. The C++ part exposes the interface for planning, which is then used by Python.

The C++ part uses OMPL to do path planning. The constraint and state validity checker are both defined in C++ part.

The Python part is used to store the sampled paths to a HDF5 file. Each path is a separate dataset with name "path[num]". There is also a dataset called "collision" that contains the collided configs.

Dependencies

• Python: 3.5/3.7
• C++: 11
• OMPL: 1.4.2
• Boost: 1.58
• HDF5: 1.8.16
• YAML-CPP

Steps for evaluating the MPNet algorithm

1. Train MPNet:
2. Generate several environments with GenerateBrickConfig.cpp.
3. Generate training samples: 1. Generate 10000 pairs of starts and goals in these environments with GenerateTrainingStartAndGoal.cpp. 1. Generate paths for these starts and goals with GenerateTrainingPath.cpp. 1. Clean up training samples with preprocess_path.py.
4. Generate voxels: 1. Generate point clouds for these environments with GeneratePointCloud.cpp. 1. Convert point clouds to voxels with convert_point_cloud_to_voxel.py.
5. Train with main.py.
6. Test MPNet:
7. Generate embedded voxel with convert_voxel_to_embedded_voxel.py.
8. Generate 1000 pairs of testing starts and goals with GenerateTestingStartAndGoal.cpp.
9. Evaluate RRTConnect's performance with TestRRTConnect.cpp.
10. Evaluate CoMPNet's performance with TestMPNet.cpp.